Abstract: In order to solve the problem that the accuracy of image registration is not high due to the large number of iterations of RANSAC algorithm, an improved RANSAC image registration algorithm is proposed. First, the reference image and the image to be registered are NSCT transformed into low frequency subband and high frequency subband. Then this paper uses the vector included angle algorithm and Structural Similarity（SSIM） to extract the edge feature points of the high frequency subband, and uses the SIFT algorithm for the low frequency subband and sets the appropriate distance threshold to extract the feature points. Finally, the improved RANSAC algorithm is used to improve the matching of feature points, and the matching points are selected to achieve image registration. The experimental results show that the proposed algorithm can effectively find more pairs of matching points and accurately remove false matching points, which obviously improves the registration accuracy.

Key words: Scale-Invariant Feature Transform（SIFT）, Random Sample Consensus（RANSAC）, image registration, Nonsubsampled Contourlet（NSCT） transformation, feature point

摘要： 为解决RANSAC算法迭代次数过多导致图像配准精确率不高的问题，提出了一种改进的RANSAC图像配准算法。首先将参考图像和待配准图像进行NSCT变换分解成低频子带和高频子带。然后对高频子带运用矢量夹角算法和结构相似性（SSIM）来提取图像边缘特征点，对低频子带运用SIFT算法并设定合适的距离阈值来提取特征点。最后利用改进的RANSAC算法提高特征点匹配精度，选择出精匹配点对，实现图像配准。实验结果表明，该算法能有效地找到较多的匹配点对，准确地去除误匹配点对，明显地提高了配准精确度。

关键词: 尺度不变特征变换（SIFT）, 随机抽样一致性（RANSAC）, 图像配准, 非下采样轮廓波（NSCT）变换, 特征点